第 17 章 利用文本挖掘技术分析文献摘要

library(tidyverse)
library(here)
library(fs)
library(purrr)
library(tidytext)
library(widyr)
library(tidygraph)
library(ggraph)

本章基于R语言文本挖掘技术，分析文献摘要。具体做法是，利用R语言tidyverse、tidytext、widyr、tidygraph、ggraph等宏包分析我校文献(Web of Science)摘要的文本信息⁶。

17.1 数据导入

为了研究的可重复性，我列出了数据获取步骤： - 打开https://www.webofknowledge.com/，进入核心合集 - 输入学校全名：比如 Sichuan Normal University - 选择“机构扩展”检索 - 选择时间范围：“2009-2018年” - 选择“SCI/SSCI/A&HCI” - 点击检索 - 文档类型精炼：”Article + Review “ - 一次显示最多 50 条，一次下载最多 500 条 - 选择“其他类型下载” + “全记录与引用的参考文献” + “win UTF” - 依此下载保存

我们共获取了 1988 条文献题录数据。

read_plus <- function(flnm) {
        read_tsv(flnm, quote = "", col_names = TRUE) %>% 
        #or
        #read_delim(flnm,  delim="\t" , quote = "", col_names = TRUE) %>% 
        #select(AU, AF, SO, DE, C1, RP, FU, CR, TC, SN, PY, UT) %>% 
        select(AB, SN, UT) #%>% 
        # mutate(University = flnm %>%                # 加入了学校名
        #                     str_split("/", 7, simplify = TRUE) %>%
        #                     .[, 6] %>% 
        #                     str_sub(start = 4)
        #        ) 
}

tbl <- here("data", "newdata") %>% 
  dir_ls(regexp = "*.txt", recursive = TRUE) %>%  
  map_dfr(~read_plus(.))
tbl

esi_plus_cas_IF_set <-  read_rds(here("data", "esiJournalsList",  "esi_plus_cas_IF_set.rds"))

complete_set <- tbl %>% 
  left_join(esi_plus_cas_IF_set, by = c("SN" = "ISSN") ) %>% 
              select(category = Category_ESI_cn, abstract = AB, pubs = UT) #%>%
              # rename(ISSN = SN) %>% 

17.2 数据规整

complete_set

数据整理和文本分词

text_df <- complete_set %>% 
  filter(!is.na(abstract)) %>% 
  unnest_tokens(output = grams, input = abstract, token = "ngrams", n = 2)
text_df

过滤无用词汇

text_filted <- text_df %>% 
  separate(grams, into = c("word1", "word2"), sep = " ") %>% 
  filter(!word1 %in% stop_words$word) %>% 
  filter(!word2 %in% stop_words$word) 
text_filted 

text_unite <- text_filted %>% 
  unite(grams, word1, word2, sep = " ")
text_unite

17.3 计算tf_idf

text_tf_idf <- text_unite %>% 
  count(pubs, grams) %>% 
  bind_tf_idf(pubs, grams, n) %>% 
  arrange(desc(tf_idf))
text_tf_idf

text_tf_idf %>% group_by(pubs) %>% 
  filter(max(tf_idf) > 9.1)  %>% 
  #dplyr::distinct(pubs)
  ggplot(aes(x = fct_reorder(grams, tf_idf), y = tf_idf, fill = pubs)) +
  geom_col(show.legend = FALSE) +
  coord_flip() +
  facet_wrap(vars(pubs), ncol = 2, scales = "free")

17.4 文本相似性

\[ \text{similarity} = \cos ( \theta ) = \frac { \mathbf { A } \cdot \mathbf { B } } { \| \mathbf { A } \| \| \mathbf { B } \| } = \frac { \sum _ { i = 1 } ^ { n } A _ { i } B _ { i } } { \sqrt { \sum _ { i = 1 } ^ { n } A _ { i } ^ { 2 } } \sqrt { \sum _ { i = 1 } ^ { n } B _ { i } ^ { 2 } } } \]

text_tf_idf %>% 
  pairwise_similarity(pubs, grams, tf_idf, upper = FALSE, sort = TRUE)

17.5 关联词汇

前面我们计算了过滤词汇text_filted，我们现在研究这些词汇之间的关联

text_count <- text_filted %>% 
  count(category, word1, word2, sort = TRUE) %>% 
  select(word1, word2, n, category)  %>% 
  arrange(category)
text_count

text_count %>% 
  filter(n  > 20) %>% 
  as_tbl_graph() 

## # A tbl_graph: 124 nodes and 94 edges
## #
## # A directed multigraph with 46 components
## #
## # Node Data: 124 x 1 (active)
##   name         
##   <chr>        
## 1 piezoelectric
## 2 rights       
## 3 lead         
## 4 elsevier     
## 5 sintering    
## 6 phase        
## # ... with 118 more rows
## #
## # Edge Data: 94 x 4
##    from    to     n category
##   <int> <int> <int> <chr>   
## 1     1    74    65 材料科学
## 2     2    75    58 材料科学
## 3     3    52    45 材料科学
## # ... with 91 more rows

text_count %>% 
  filter(n  > 20) %>% 
  as_tbl_graph() %>% 
  ggraph(layout = "fr") +
  geom_node_point() +
  geom_edge_link(aes(color = category, edge_width = n)) +
  geom_node_text(aes(label = name), repel = TRUE) +
  #facet_wrap(vars(category), ncol = 3, scales = "free")
  facet_edges(vars(category), ncol = 3, scales = "free")

17.6 下一步工作

• 数据量很多，需要精炼，从而提前有用的关键信息
• 还没想好

---

6. http://tidytextmining.com/ngrams.html↩